Research On Dynamic Weighting Strategies And Adaptive Learning Of Data Stream

Data streams are ubiquitous in our daily life.Data streams are generated rapidly at all times,such as ATM transactions,sensor networks and stock transactions,and many other applications will generate data streams.Data stream can be regarded as a sequence of data arriving continuously over time,and it is a huge amount of dynamic data Different from ordinary data mining technology,data stream mining requires one scan,fast processing,and the model has the ability to dynamically update.Therefore,an efficient adaptive learning strategy is the focus of data stream mining research.The current data stream mining algorithms mainly face two major problems:concept drift and feature drift.The target concepts in the data stream may evolve over time.This change is called as concept drift.The correlation of features in data streams also changes over time,which makes the subset of features no longer relevant to the target concept,resulting in a special type of drift called feature drift.The concept drift and feature drift will seriously affect the classification accuracy of classifiers.At present,the processing efficiency and effect of the existing algorithm for processing concept drift are still not ideal,and most data flow algorithms do not fully consider the situation of feature drift.Therefore,this paper studies the concept drift and feature drift in the data stream,and the dynamic weighting strategy is used to adapt to the concept drift and feature drift.That is to say,the classifier has the ability of adaptive learning.The main work includes:(1)An instance dynamic weighted Bayesian classification algorithm is proposed to deal with the concept drift problem in the data stream.The newer instance has a greater impact on building the classifier,so this paper proposes a time decay function,which uses forgetting mechanism to attenuate the weights of instances in data chunks,and set a weight threshold.With each iteration,the weights of older instances will decrease until they are discarded below this threshold.Then the instance weighting module is applied to the Bayesian classifier to improve the original Bayesian classifier to adapt to the concept drift.Experiments show that the method can effectively handle the concept drift in data stream mining.(2)An feature dynamic weighted Bayesian classification algorithm is proposed to deal with the feature drift problem in the data stream.For the dynamic feature space of data stream,the method of feature weight calculation based on correlation is used.The sliding window technology is used to track the correlation trend between the feature and the target concept and other features,and then the feature weight is calculated.At the same time,a dynamic feature weighting model is established for Bayesian to reduce the influence of feature drift.The learned feature weights can improve the classification performance of Bayesian model as a whole.Experiments show that the classification accuracy of this algorithm is improved compared with other algorithms.